Why MCPs Are Making Your AI Dumber

Anthropic admits that MCPs suck, and they do suck. They eat up context. They they're unreliable at best, and they make your AI dumber because again, that additional context is causing a I think they call it um needle in a haststack uh problem or something like that. People have realized that it's a standard that needs reinvention. And who better to make that realization or to admit that realization than the very people who invented it? Anthropobbit. Theo made this video just kind of basically going over this article which raises some really interesting points. It again it just admits to the the tool calling aspect constantly having these tool calls in context whether you're um whether you're using them or not. I it's a headache to be honest. While I'm coding I'm turning MCPs on and off and then I forget to turn them on and off and you know it's just it's something you just don't want to have to manage. Just a quick side note I did learn because I'm starting to use Code Rabbit. I'm starting to build up a bit more robust um coding environment that picks up on my blind spots allows me to be a more uh impactful independent programmer. Uh, I install Code Rabbit, which is an AI code reviewer. Um, and they get you to use the CLI through Clawude Code. What would you do in terminal? Do you use any CLIs that we could Yeah, in my terminal, I just do git. You've got git installed on your computer and you you trigger that CLI with the term git. If you start that terminal command with git, that's using now the CLI and they're the arguments that follow that. Hope that makes sense. I didn't know that. So okay that makes sense because I do get status right yeah to get to understand if I have pushed or not I forget and then I do get add dot to add everything essentially and this is a great example because you can ask your um AI to push if you just say push this to git it will do get commit and and write a nice commit message and get push so it I always knew it could use a CLI because it's using the git CLI uh you can use it with Code Rabbit. Similarly, I use Taskmaster which eats up so much context. Taskmaster has a CLI which I've got installed. So now I'm just going to remove that MC. I have removed that MCP and I'm going to use the CLI. That's just my own personal thing. That's nothing to do with this article, but it's a way of me now just really um you know realizing that MCPs just suck. Okay, they they just I don't I I prefer to preserve context and not have to manage stuff. uh and have my AIO not be stupid. So, here's here is uh Anthropic going through the downsides. They're not going to completely tell you it's rubbish. They still keep their kind of one foot in the door with MCP. They still want you to use it. But what it does go on to say, and what I've noticed recently, actually, so it's it's funny how it all comes about because by the way, this was released I think on the 4th of November. Yeah, 4th of November. What I've noticed that AI is doing now is writing JavaScript files to do sort of push stuff to to the database or to do certain things. So it will write its own code and AI has always been great at writing its own code. So what what Claw Code is suggesting here is set and I've yet to understand how to actually do this. Do you ask Claude to create this MCP server? Um, but ultimately you you tell it that there is the MCPS are here as files that it can use. So you can remove the MCP and this just uses the APIs um to to access Google Drive instead of the Google Drive MCP as a bunch of files there which files don't eat up context unless it's actually going to be using them. So then the AI will write a piece of code that imports um actually this is the wrong one here. Uh it's this one. It imports the Google Drive files and then runs this bit of code and then this is this is the only context that it's using, you know, because it just runs this code and then it's gone. It's not going to be wasting up. And what they found was 150,000. This is wild, by the way. 150,000 tokens to 2,000 tokens, a time cost saving of 98%. What are you doing, Anthropic? But why have you why are you still using this MCB this standard that's that you can if if you can rejig it just a little bit you know maybe this is a skill maybe this is something like 98.7% is crazy crazy so I want to start looking into how I can incorporate what they're doing here I'll I'll read the article a little bit more deeply but and how I can inevitably create a you know a file-based MCP service or whatever if indeed Google Drive doesn't have a a CLI that I can just tap into because a CLI will just run, but ultimately you're asking your AI just to run a bit of code. Um, so yeah, they're admitting MCP is is not so great and there are other ways to use it, but in the summary here somewhere they again they kind of keep their foot in the door. MCPS are they on their way out? I think people struggle with the concept of them. I would love it if they work if they were more deterministic if they would always run. I've I've resorted to saying use the taskmaster MCP to add this as a task or use the linear MCP. You know, I've been very It used to be that you could lightly mention, you know, add this task or um what Figma like uh take the design or something like that and it will just know to use the MCP. But it's just too hit or miss. you might as well just use use the Figma MCP to you know get that node or whatever it is. They are it is quite nice that you can again just ask something to do something in natural language. That's great. It really is. But there needs to be a way to do it without um you know having to uh constantly I don't know it's just they just don't need to be there basically. So yeah. Yeah. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter.